1. The Wrong Question
The debate about automation and employment asks whether machines will replace human workers. Optimists cite two centuries of evidence: the loom displaced weavers but created loom operators, the tractor displaced farm hands but created mechanics, the computer displaced typists but created an entire digital economy. Technology has always created more jobs than it destroyed. Why would this time be different?
The answer is not that all jobs disappear. The rate of displacement is outpacing the rate of new opportunity creation, and the gap is widening. This is not about how many jobs go. It is about how fast.
But even the rate question misses something. There is a deeper mechanism at work, one that makes the displacement self-reinforcing in a way previous technological transitions were not. I call it the Agentic Fluency Trap: the compounding cycle where humans get better at directing AI, which displaces other humans, regardless of whether the AI itself improves. The trap is that the learning curve for directing AI is itself AI-assisted, which means it steepens faster than any previous technology adoption curve in history.
2. The PlayStation Dynamic
Consider the PlayStation 3. Early titles looked primitive compared to the games released in the console’s final years. The hardware was the same silicon throughout. Developers learned to extract more from a fixed capability. The visual quality, the physics simulations, the complexity of the worlds all improved dramatically without a single transistor being added to the machine.
The same dynamic is playing out with large language models, except the thing being optimised is not frame rates but headcount.
Consider a team of 10 employees, A through J, doing similar knowledge work. Employee A starts using existing AI tools to automate part of their workflow. Not a breakthrough model. The same large language model that has been available for months. A builds a prompt template, integrates it into their daily process, and discovers it handles a chunk of work that used to take hours. A is now noticeably more productive. Employee J, the least productive member of the team, is made redundant. The company has the same output from 9 people.
A continues to refine the process. A uses the AI itself to improve the tooling. The AI helps A write better prompts, build more tailored workflows, automate the integration between systems. The model has not improved. A’s ability to exploit it has, assisted by the model itself. A shares the refined process with B. Now both A and B are significantly more productive. H and I are redundant. The company has the same output from 7 people.
A and B keep refining. G goes. F goes. E goes. The team is now A, B, C and D doing the work of 10. The AI model is identical to the one A started with. What changed is the accumulated human ingenuity applied to exploiting it, with the AI assisting at every step.
This is the trap. The model does not need to improve. The humans using it will keep getting better, keep building better tooling, keep finding new ways to do more with less. And they will do it with the help of the model itself. Every iteration of tooling improvement is AI-assisted, which means it happens faster than the previous iteration. The learning curve compounds.
3. From Prompts to Architecture: Three Phases of Fluency
The fluency trap operates through three distinct phases of human skill development. Each phase represents a step change in how effectively a person can direct AI, and each phase is learned with the help of the AI itself.
3.1 Prompt Engineering
This is where everyone starts. Learning to talk to the model. Crafting questions, structuring requests, providing examples, iterating on phrasing until the output improves. The web UI era. It is useful and it is where roughly 90% of AI users remain today.
Prompt engineering produces linear improvements. A better prompt gets a better answer. But the ceiling is low, because the model’s reliability degrades with task complexity. A single prompt can produce a function. It cannot reliably produce a system.
3.2 Tool Engineering
The step change comes when a practitioner stops asking the model to do everything and starts building deterministic scaffolding around it. The model reasons. Scripts execute. The key insight is separation of concerns: push reliability into deterministic code (tools that do one job, do it fast, and never hallucinate) while pushing flexibility and judgment into the model (deciding which tools to use, in what order, and how to handle edge cases).
This matters for displacement because of compound reliability. If a model is 90% accurate per step, five chained steps give you roughly 59% accuracy. That is not useful. But if four of those five steps are deterministic scripts that are 100% reliable, and only the judgment step uses the model, the compound accuracy jumps to 90%. The model has not improved. The architecture around it has eliminated its failure modes.
Tool engineering is where the practitioner moves from using AI as an answering machine to using it as a manager that delegates to reliable subordinates. The productivity multiplier jumps from marginal (prompt engineering) to substantial. One person with a well-scaffolded AI workflow does the work of three to five people doing the same work manually.
3.3 Context Engineering
The third phase is the most recent and the most powerful. Context engineering is the practice of curating what the model sees at each step of inference. It treats the model’s attention as a finite resource with diminishing marginal returns (which, architecturally, it is) and optimises for the smallest possible set of high-signal tokens that maximise the likelihood of the desired outcome.
In practice, this includes:
- Memory systems. The model writes structured notes to persistent storage and reads them back into context at later steps. This gives an agent continuity across tasks that exceed the context window. Without memory, an agent forgets. With memory, it builds tacit knowledge over time.
- Context compaction. When a conversation approaches the context limit, the model summarises and compresses older context, preserving critical details while discarding redundant output. This allows an agent to work for hours without degrading.
- Sub-agent architectures. Multiple specialised agents work in parallel, each with a clean context window focused on a specific subtask. Each agent explores extensively (tens of thousands of tokens) but returns only a condensed summary to the lead agent. The lead agent never sees the detail. It synthesises the results.
- Progressive disclosure. Instead of loading all relevant data up front, agents maintain lightweight references (file paths, stored queries, URLs) and retrieve data just in time. Like a human who does not memorise an entire library but knows where to look.
Context engineering is what turns a capable model into a capable worker. It is the difference between a model that can answer a question and a model that can run a project. And like the previous phases, the model helps you learn it. You use the AI to build the memory systems that make the AI more effective. You use the AI to design the agent architectures that let the AI manage itself. The recursion is the trap.
3.4 The Mirror in Software Engineering
These three phases map precisely onto a progression in how software is built with AI:
| Level | Role Analogy | How AI is Used | What Gets Reviewed | Displacement |
|---|---|---|---|---|
| L1: Answerbot | Student | Web UI, specific Q&A, build a function | Line by line | None (most users are here) |
| L2: Code Assistant | Junior Dev | IDE plugin, assisted coding, autocomplete | Commit level | Low (speeds up individual) |
| L3: Senior Dev | Senior Dev | Agentic coding, builds features autonomously | Pull request level | Medium (1 does work of 2-3) |
| L4: Dev Manager | Engineering Manager | Agent orchestration, multi-PR milestones | Outcomes, not code | High (1 does work of a team) |
| L5: Product Manager | Product Lead | Multi-workstream agentic build, testing by separate agents | Final product functionality | Extreme (1 person, full product) |
L1 and L2 are prompt engineering. L3 is tool engineering (the agent has tools, uses them, but a human reviews the output at the level of a pull request). L4 and L5 are context engineering (the human manages what the agents know, not what they do; review shifts from code to outcomes).
The transition from L3 to L4 is also a shift in how the human relates to the agent. At L3 and below, the human is in the loop: every deliverable passes through them for approval before the agent can proceed. At L4, the human moves on to the loop: the agent executes autonomously and the human monitors outcomes, intervening only on exception.
This is what unlocks the full time horizon of a capable model. An agent that can work for 12 hours autonomously but must pause for human approval every 30 minutes wastes 95% of its capability.
The displacement implications of this shift are severe. A human in the loop can gate one agent at a time. A human on the loop can oversee multiple agentic workflows concurrently. The bottleneck is no longer “can the agent do the work?” It is “how many agents can one human monitor?”
At L4, one person manages a team of agents the way an engineering manager manages a team of developers: setting goals, reviewing outcomes, handling escalations. At L5, the agents manage each other (testing agents validate the output of building agents) and the human reviews only the final product. The displacement ratio is no longer 1:3 or 1:5. It is 1:N, where N is bounded only by the human’s capacity to define goals and evaluate results.
Most organisations today are at L1 or L2. The displacement pressure comes from the minority at L3 and above spreading their practices. And every level on this ladder is achievable on current models. L5 is not theoretical. It is being done today, with existing tools, by people who have climbed the fluency curve. The models already shipped are sufficient. The humans have not caught up yet. When they do, the displacement accelerates.
4. Beyond Exponential: The Shape of the Curve
The non-profit research organisation METR (Model Evaluation and Threat Research) maintains the most rigorous independent measurement of AI agent capability. Their metric is the time horizon: the length of task (measured by how long it takes a human professional) that an AI agent can complete autonomously with a given probability of success. They measure this across a diverse suite of software and reasoning tasks and publish updated results as new models are evaluated.
Plot the time horizon on a logarithmic scale and you expect a straight line if improvement is exponential. The METR graph does not show a straight line. It curves upward.
The doubling time (how long it takes for the autonomous task length to double) has been shrinking:
| Period | Doubling Time |
|---|---|
| 2019-2026 (all time) | 188 days (~6 months) |
| 2023 onwards | 129 days (~4 months) |
| 2024 onwards | 89 days (~3 months) |
A shrinking doubling time means the rate of improvement is itself accelerating. On a log scale, exponential growth is a straight line. Super-exponential growth curves upward. That is what the data shows.
The raw numbers tell the same story. In February 2025, the frontier model (Claude 3.7 Sonnet) had a 50% time horizon of 60 minutes: it could autonomously complete tasks that take a skilled human one hour, half the time. One year later, in February 2026, Claude Opus 4.6 has a 50% time horizon of 719 minutes: nearly 12 hours. The upper confidence bound extends to 3,950 minutes (66 hours, or roughly 8 working days). That upper bound is wide because METR’s benchmark suite is running out of tasks long enough to properly measure the model. The ceiling of the test is lower than the ceiling of the model.
Twelve hours of autonomous work at 50% reliability. That is not a tool assisting a worker. That is the worker. A single AI agent, with no human intervention, completing tasks that would take a skilled professional an entire working day (and then some). This was measured independently by a research non-profit, not claimed by the model’s developer.
The progression through the top of the METR chart tells the story of acceleration:
| Model | Release | 50% Time Horizon | Jump |
|---|---|---|---|
| Claude 3.7 Sonnet | Feb 2025 | 60 min (1h) | |
| o3 | Apr 2025 | 120 min (2h) | 2x in 2 months |
| Claude Opus 4 | May 2025 | 100 min (1.7h) | |
| GPT-5 | Aug 2025 | 203 min (3.4h) | |
| Opus 4.5 | Nov 2025 | 293 min (4.9h) | |
| GPT-5.2 | Dec 2025 | 352 min (5.9h) | |
| Claude Opus 4.6 | Feb 2026 | 719 min (12h) | 2.5x in 73 days |
From 1 hour to 12 hours in one year. A 12x increase. And this measures raw model capability on a standardised benchmark with a simple agent scaffold. It does not capture what happens when a fluent human adds tool engineering, memory systems and context architecture on top.
5. The Paradox That Proves the Trap
In July 2025, METR published a randomised controlled trial measuring the impact of AI tools on experienced open-source developer productivity. The finding was counterintuitive: developers using AI took 19% longer to complete tasks than developers working without it. AI made them slower.
This result was widely cited as evidence that AI is overhyped. It is nothing of the sort. It is evidence of the fluency trap in action.
The study used early-2025 tools (Cursor Pro with Claude 3.5 and 3.7 Sonnet). The developers used chat, autocomplete and IDE agent mode. That is L1-L2 on the fluency ladder. Their average prior AI tool experience was “dozens of hours.” The tasks averaged two hours.
The study explicitly noted its limitations: it did not capture what happens with better scaffolding, longer learning curves, or different usage patterns. It measured developers near the bottom of the fluency curve.
Now compare: at the time of that study (mid-2025), the frontier model had a time horizon of roughly 100-200 minutes. By February 2026, Claude Opus 4.6 has a time horizon of 719 minutes. The model has outrun the methodology that said it was not helping.
The fluency trap predicts exactly this pattern. L1-L2 users are slower because they have not learned to direct the model effectively. They fight with autocomplete. They rephrase prompts. They review AI output line-by-line and find it quicker to write the code themselves. But L3+ users, who have built deterministic scaffolding and learned to manage context, are dramatically faster. The METR study captured the left side of the fluency curve and accidentally demonstrated why the trap is invisible to those who have not climbed it.
The developers believed AI had sped them up by 20% when it had actually slowed them by 19%. The perception gap matters. Organisations adopting AI at L1-L2 will see modest or negative productivity gains and conclude the technology is not yet disruptive. Meanwhile, a smaller number of practitioners at L3+ are quietly doing the work of teams. The aggregates conceal the distribution.
6. What Drives the Trap
Three mutually reinforcing dynamics explain why fluency compounds faster than any previous technology adoption:
6.1 Tool Maturation
Deterministic scaffolding is becoming standardised. Open protocols like the Model Context Protocol allow agents to interact with external systems through a common interface. Architectural patterns for separating probabilistic reasoning from deterministic execution are emerging across the field: the model decides what to do, scripts do it, tools handle API calls and data processing, and the model never touches the operations directly. This eliminates the compound reliability decay that makes raw model output unreliable over long task chains.
These patterns are not proprietary or difficult to learn. They are open source, well documented and actively taught by the model itself. Ask any frontier model how to build a reliable agent architecture and it will describe the same separation of concerns. The tooling is standardising faster than previous technology stacks because the technology helps you learn it.
6.2 Context Engineering Maturation
The practice of curating what a model sees is becoming a recognised engineering discipline. Compaction (summarising older context to free space), structured memory (persistent note-taking outside the context window), sub-agent delegation (parallel specialised workers returning summaries) and progressive disclosure (just-in-time data retrieval) are now documented best practices from the model developers themselves. These techniques multiply a model’s effective capability without the model changing. A 12-hour time horizon with naive scaffolding becomes substantially longer with competent context management.
6.3 Self-Reinforcing Learning
You use the AI to build better tooling for the AI. The AI helps you write better prompts, build better scaffolds, design better memory systems. Each iteration makes the next iteration easier and faster. A practitioner at L3 uses the model to build the tools that move them to L4. At L4, they use agent orchestration to build the multi-agent systems that enable L5.
This is the core of the trap. The learning curve for exploiting AI is itself AI-assisted. Previous technologies had human-speed learning curves: it took years to become a proficient programmer, decades to master a trade. The fluency curve for AI tooling is compressed because the teacher and the tool are the same thing.
7. Invisible in the Averages
The displacement hides in aggregate statistics. When Employee A does the work of Employees A through J, the company’s headcount drops but its revenue stays the same or increases. Employee A is more valuable and commands higher pay. Average earnings rise. Corporate profits increase (lower wage bill, same output). GDP holds up. The headline numbers look healthy.
What the aggregates conceal is a widening distribution. Fewer workers earning more. More people earning nothing or much less. Average earnings increase while the median hollows out. Corporate profits rise while benefit claims grow. Every headline indicator says the economy is performing while the lived reality for a growing number of people says the opposite.
The Office for National Statistics reported in February 2026 that 891,000 people aged 18 to 24 in the UK were not in education, employment or training. That is 15.2% of the age group, and the figure is rising quarter on quarter. Goldman Sachs Research published in August 2025 that unemployment among 20- to 30-year-olds in tech-exposed occupations had risen by almost 3 percentage points since the start of 2025. BT announced plans to shed up to 55,000 workers, citing AI. The OECD reported in 2024 that generative AI exposure is greatest for high-skilled workers and women, reversing the pattern of previous automation waves.
None of this shows up as a fiscal emergency yet because the aggregates mask the distribution. The crisis does not announce itself. It arrives the day the aggregate tips, when the displaced outnumber the remaining productive workers enough to make the averages turn. By then, the structural damage is years old.
8. The Fiscal Implication
The UK tax system’s primary tools for managing aggregate demand depend on employment. Income Tax (£329B) and National Insurance (£205B) together withdraw 42% of government spending from the economy. Both are levied on wages. As automation displaces workers, less purchasing power is withdrawn while transfer payments grow. The state loses its main mechanism for managing demand without gaining a replacement.
This is not a fiscal crisis in the orthodox sense (the government running out of money). A sovereign currency issuer cannot run out of its own currency. The crisis is one of inflation management: the state’s ability to withdraw purchasing power from the economy weakens precisely as the need for public provision (income support, retraining, services for the displaced) grows. If the state spends more without withdrawing more, the result is inflationary pressure. If it refuses to spend, the result is destitution. Neither is acceptable.
The rate of displacement matters for the policy response. If displacement is slow and cyclical, a labour-based buffer stock (a job guarantee at a fixed wage, absorbing displaced workers and releasing them as private demand recovers) could, in principle, manage the transition. But the fluency trap produces secular, accelerating displacement. The buffer stock does not shrink because private sector demand for the displaced workers does not recover. The jobs are not coming back. The model already does them.
A job guarantee anchors prices by setting a fixed wage for the buffer stock. But if the binding constraint on production shifts from labour to energy and compute (which is the trajectory the METR data describes), then a wage anchor stabilises a price that is becoming irrelevant. You do not need a wage anchor when wages are a diminishing share of production costs. You need a mechanism that captures the actual inputs to automated production: energy and data.
The Sovereign Energy and Bandwidth Excise (SEBE) taxes the physical infrastructure of automated production: energy consumption (kWh) at the point of load and cross-border commercial data (TB) at the border. Every prompt, every API call, every model inference consumes electricity at a data centre. That consumption is physical, metered and unavoidable. SEBE captures it in real time, regardless of whether HMRC has noticed the headcount reduction. SEBE withdraws £34-46 billion in purchasing power at launch (2030, 2026 prices), growing to £93 billion by 2040 as automation scales. All rates are CPI-indexed mechanically. No commission, no political discretion, no annual review. The meter reads what it reads.
Income tax sees the symptom of displacement, and it sees it late because rising average earnings and corporate profits mask the underlying erosion. SEBE sees the cause, from day one, at the point of energy consumption. Income tax shrinks as the problem grows. SEBE grows with it.
9. The Choice
The tools are already sufficient. The displacement is already underway. The models do not need to improve. The humans using them will keep getting better, keep building better tooling, keep finding new ways to do more with less. That is what humans do. It is also what makes the current fiscal architecture unsustainable.
The METR time horizon graph curves upward on a log scale. That is super-exponential improvement. The fluency trap compounds on top: human tooling improvements, AI-assisted and accelerating, multiply whatever the raw model can do. Any policy framework built on the assumption that displacement will be slow, cyclical or decades away is already behind the curve.
The choice is not between automation and employment. It is between a tax system that watches the curve steepen and one that captures the energy powering it.
The full SEBE summary is available at Tax Robots, Fund People.
Jason Huxley is an infrastructure and automation engineer with 30 years’ experience in large-scale enterprise systems. He is a member of the Green Party of England and Wales. The full SEBE proposal is open-source at github.com/djarid/SEBE under CC-BY 4.0.
Copyright 2026 Jason Huxley. Licensed under CC-BY 4.0.